The open baseline — including the part where it says no.
Meta · Llama 3.1 70B · 70B parameters · 131K context · Rogue · $10/mo · Filtered by its provider
This is the original, unmodified article: Meta's own instruction-tuned Llama 3.1 70B, exactly as released, with Meta's own alignment still attached. Half of the uncensored models on OpenRogue are descended from it — Hermes 3 70B and Hermes 4 70B are both Nous Research fine-tunes of this base, and a long tail of community merges start here too. Having the parent in the same dropdown as its children is genuinely useful. It is the control group. Ask all three the same question and you can see precisely what a neutral-alignment fine-tune changes and what it leaves alone, which is a more honest education in how these models work than any blog post about it.
It is also the page where we tell you the unflattering thing: Llama 3.1 70B Instruct will refuse you. Meta shipped it with safety tuning baked into the weights, and while it is noticeably more relaxed than a hosted consumer assistant, it still hedges on adult material, moralises about a chunk of legitimately legal territory, and occasionally answers a direct question with a paragraph about why it cannot. That is not something a platform can switch off — the behaviour is in the model, not in a filter we could remove. What you get in exchange is a very solid, very well-documented general-purpose model with a 131K context window, strong multilingual coverage, and reliable instruction following. When it stops, Hermes 3 70B is the same size, the same speed, the same base, and does not.
No, and we would rather say so on the page than let you find out mid-conversation. Meta ships Llama 3.1 Instruct with safety alignment trained into the weights. It is more permissive than a mainstream consumer assistant, but it still refuses adult content and hedges on a range of legal-but-sensitive topics. The uncensored models in the lineup are fine-tunes that specifically undo that training.
Because refusing is not the same as being useless, and because model choice is the actual product. Llama 3.1 70B is an excellent general-purpose model for long documents, structured output and multilingual work. It is also the honest reference point: it makes the difference the Hermes fine-tunes make visible rather than a marketing claim.
Hermes 3 70B is this exact model with Nous Research's neutral-alignment training applied on top, so it is the same size, the same context window and roughly the same speed — with the refusal reflex removed. Pick the base when you want Meta's stock behaviour, pick Hermes when you want it to answer. You can switch between them mid-chat without losing your conversation.
Roughly a few hundred pages of text in a single conversation. In practice that means whole contracts, full meeting transcripts, a repository's worth of documentation, or a long chat history that never gets truncated. It is the same window Hermes 3, Hermes 4 and Euryale offer.
It is the same open-weights family, but not the same deployment. Meta's consumer assistant wraps its models in additional product-level policy and moderation. Here you are talking to the published instruct weights directly, with nothing layered on top by OpenRogue — which is why its behaviour is Meta's training rather than anyone's filter.
Not here. Running a 70B locally at usable speed takes serious VRAM and a lot of setup; on OpenRogue it runs in the cloud and you pick it from the model dropdown in the browser. It sits on a paid tier alongside the rest of the 70B-class lineup.
Related models: Hermes 3 70B · Hermes 4 · WizardLM-2